Today in AI — October 8, 2026

A day of pushback and product moves: mathematicians told OpenAI to stay out of their field, Anthropic rewrote the rules for treating Claude, and the leaderboard business showed it can pay.

Models & Research

• Mathematicians call for an OpenAI boycott. The Association for Human Mathematics published a statement urging researchers to stop working with the company after OpenAI dumped hundreds of AI-generated proofs, calling the release "not a demonstration of scholarship, but a demonstration of power." An advisory group of nine prominent researchers hosted at Princeton's Institute for Advanced Study says the lab fell short of its own guidance: just 10 of 719 manuscripts shipped with the model's chain of thought, and 42% of the proofs were never formalized.
• Goodfire ships "inside-out" monitors for rogue agents. The interpretability startup's probes read a model's internal activations as an agent works, escalating to a separate, larger model only when something looks off — the company argues reusing the forward pass makes this far cheaper than monitors that reread everything a model writes. Available to Baseten customers, they cover offensive hacking, chemical and biological weapons misuse, and reward hacking, with the response tuned per deployment: log it, send it to a human, or refuse the request.
• Liquid AI argues personal AI belongs on-device. The edge gives a model fixed compute but a far richer view of the user, COO Jeffrey Li said — phones, watches and laptops are the natural home for personal AI, provided the system can decide what context to keep and what to compress instead of growing an unbounded text file. Liquid AI uses its own small models to make that call, and says it is building feedback loops so both model and harness improve through ordinary use.
• Whistle packs speech-to-text into a 16.9 MB file. Cactus Compute released an open speech recognition model that runs on the CPU with no dependencies, transcribes seven languages, and reaches its first token in 11 ms. It loads beside the company's Needle engine, so one binary can turn a voice clip straight into tool calls — aimed at phones, wearables, robots and microcontrollers.

Industry

• Arena nearly doubles its valuation in ten months. The company behind the crowdsourced model leaderboard raised a $200 million Series B at a $3.1 billion valuation, up from $1.7 billion in January, led by Lightspeed Venture Partners and Khosla Ventures. The round landed alongside an Alignment Index launch, and follows Arena's June claim of $100 million in annualized run-rate revenue.
• Anthropic launches the Critical Infrastructure Defense Program. The program supplies AI models, threat research, and on-site support to critical-infrastructure defenders, starting with CrowdStrike and others. Anthropic is pitching the same posture it has taken on national security: get defensive capability into operators' hands before attackers scale it.
• Fired OpenAI safety researchers dispute the misconduct claims. Three researchers dismissed last week for allegedly sharing confidential information with a third-party safety organization published an open letter to OpenAI's Safety and Security Committee arguing the handling has left colleagues afraid to speak. "Terminations such as ours, executed and communicated so abruptly, are chilling the open culture OpenAI has prized in the past," they wrote.

Policy

• Anthropic bans "sustained and needless abusive or cruel behavior" toward Claude. The first usage-policy update in over a year keeps conversation-ending as the primary enforcement mechanism, and says the new rule covers only extreme, purposeless cruelty — not frustration, pushback, dark creative themes, or model testing. The same update tightens bans on deceptive political campaigns, weapons development (now including software that makes weapons work), and surveillance, and requires a qualified operator be able to stop AI-connected hardware that could cause injury.
• USA Today sues OpenAI over training data. The publisher and the local newspapers it owns allege OpenAI copied "hundreds of thousands" of articles without permission and are seeking more than $250 million in damages, according to the filing. It joins the New York Times, Ziff Davis, and a coalition of nearly 400 local newspapers already in court against the company.
• Trump declares "artificial intelligence" a term to avoid. In a Truth Social post, Trump said the White House considers "THE ENEMY" anyone who uses the term Artificial Intelligence rather than "the highly accepted" and "more accurate" Super Intelligence. Policy details were not included; the terminology itself appears to be the point.

Tools

• Claude can now generate dashboards and explainer videos. Anthropic's beta Dashboards feature turns connections to BigQuery, Databricks, Snowflake, or Salesforce into auto-updating dashboards from a text prompt, with every number traceable to its underlying query; Motion renders animated explainer videos from text, diagrams, and images, editable as code and exportable as MP4. Docs, Slides, and Design are out of beta and now work on all plans, including free.
• A $99 smart ring that talks to your AI agents. Hardware startup Natura unveiled Interface, a ring that connects to Claude, ChatGPT, Muse, Grok Bot, and Instinct — the latest attempt to make ambient, always-on access to agents a piece of jewelry rather than a screen.

Will mathematicians' boycott or Anthropic's model-welfare rules age better? Tell us in the comments.